Sei sulla pagina 1di 14

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS 1

Multilevel Power Optimization of


Pipelined A/D Converters
Jintae Kim, Member, IEEE, Sotirios Limotyrakis, Member, IEEE, and Chih-Kong Ken Yang, Senior Member, IEEE

Abstract—Power dissipation of analog and mixed-signal cir- the signal current, leading to a low-power pipeline stage imple-
cuits has emerged as a critical design constraint in today’s mentation. At architecture-level, optimizing per-stage resolu-
VLSI systems. This paper presents a multilevel design opti- tion and noise contribution have been commonly used [5]–[10].
mization approach for reducing the power dissipation of a
pipelined analog-to-digital converter (ADC). At the circuit-level,
Various calibration techniques in both analog and digital do-
device-types and supply-voltages are jointly optimized for the main [11]–[13] can compensate the inaccuracy of raw analog
residue amplifier of a pipeline stage to minimize power. At the signal processing, relaxing the precision requirement of analog
architecture-level, the nonlinearity contribution from stage gain circuits and thereby achieving low-power dissipation. In many
error is optimally distributed to further minimize combined power cases, these techniques are not mutually exclusive.
dissipation. The optimizations take advantage of an analytical This work demonstrates an analytical approach for reducing
optimization method based on geometric programming for a
quantitative tradeoff analysis. All of the proposed power opti-
the power dissipation, which in principle can supplement most
mizations are applied to the design of a two-way interleaved 8-bit of the existing techniques. We exploit a fundamental observa-
320 MS/s pipelined ADC in 90-nm CMOS technology. Measured tion that a broader design variable space, when a global op-
performance from a prototype chip shows 7.30-bit of ENOB at timum is found, improves the power-performance envelope of
Nyquist input frequency with DNL of 0.35 +
0.45 LSB and INL
of 0.72 + 0.89 LSB, while dissipating 12.77 mW from 2.1 V/1.2
the system. Two such design approaches are proposed in this
work. At the circuit-level, we propose to jointly optimize de-
V supplies. The achieved conversion efficiency is 253fJ/conv-step.
vice types and supply voltages of the residue amplifier to imple-
Index Terms—Analog circuit, data converter, low-power design, ment low-power pipeline stage. At the architecture-level, contri-
optimization, pipelined analog-to-digital converter (ADC). butions of per-stage gain error to the total linearity degradation
are optimized in addition to optimizing per-stage resolution and
noise contributions for minimizing total power dissipation.
I. INTRODUCTION Adding degree of freedoms to the variable space increases the
design complexity, often times exponentially. This work takes

P OWER dissipation has emerged as the most critical design


constraint in modern VLSI systems [1]. While tradition-
ally more critical for battery-operated applications, power dis-
advantage of a design optimization method based on geometric
programming (GP) [14], a special type of convex optimization
problem, to address the increased design complexity. An equa-
sipation has become a universally important design metric due tion-based pipelined ADC model compatible with GP is de-
to growing power density in today’s VLSI as well as the lim- veloped to enable an efficient tradeoff analysis for proposed
ited cooling and power delivery capacity of the package. With power optimizations. This approach yields an optimized tran-
ongoing trend of increasing integration of synthesized digital sistor-level design when applied to a predefined circuit topology
circuit and custom analog circuit, power dissipation from both and process technology, an additional benefit not available in
types of circuit blocks must be considered in order to achieve hand analysis.
an overall low power dissipation of VLSI systems. The focus of The rest of this paper is organized as follows. Section II
this paper is to demonstrate a design approach that minimizes presents an overview of the pipelined ADC model that is used
power dissipation of analog and mixed-signal buliding blocks. as a testbench of the analysis in this work. Section III explores
Pipelined analog-to-digital converter (ADC), a popular ADC the joint-optimization of device-type and supply voltages to
architecture for medium-to-high resolutions (8–12 bits), is one designing low-power pipeline ADC stage. Section IV de-
such building blocks that are commonly used in modern mixed- scribes the optimization of nonlinearity contribution from stage
signal ICs. To minimize the power dissipation, a number of ap- gain errors at the architecture level. The tradeoff analyses in
proaches have been proposed from a wide range of perspec- Sections III and IV take advantage of the ADC model described
tives. Circuit-level techniques such as opamp sharing [2], [3] in Section II to quantitatively explore corresponding design
and comparator-based charge transfer [4] aim to efficiently use tradeoffs. Proposed power optimizations are applied to the
design of an 8-bit 320 MS/s pipelined ADC in 90-nm CMOS
Manuscript received September 06, 2009; revised January 11, 2010. This technology. Sections V and VI describe the chip implemen-
work was supported by DARPA-TEAM. tation details and measured performance from a prototype
J. Kim was with the Electrical Engineering Department, University of Cal-
ifornia, Los Angeles, CA 90095 USA. He is now with Agilent Technologies, chip. Section VII concludes the paper by summarizing main
Santa Clara, CA 95051 USA (e-mail: jintae.kim@ieee.org). contributions. In the three Appendies, the GP-based design
S. Limotyrakis is with Atheros Communication, Inc, Santa Clara, CA 95051 methodology is briefly reviewed and the circuit-level pipeline
USA.
C.-K. K. Yang is with the Electrical Engineering Department, University of
stage model is described in more detail. Additionally, a mod-
California, Los Angeles, CA 90095 USA. eling technique that can transform a nonconvex inequality to a
Digital Object Identifier 10.1109/TVLSI.2010.2041077 set of GP-compatible monomial inequalities is presented.
1063-8210/$26.00 © 2010 IEEE
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS

Fig. 3. Stage-level model.


Fig. 1. Hierarchies in pipelined ADC model.

front-end stage is a gain-of-one stage without a sub-ADC. The


backend stage is a sub-ADC without an MDAC and redundancy.
Therefore, no separate modeling is necessary for the front-end
and back-end stage.

B. Variables in the Model


The ADC model in the top-level and stage-level considers
following variables. The subscript in the variable refers to the
th stage.
• Peak-to-peak analog input swing: .
Fig. 2. Top-level model. • Stage resolution: .
• Stage total sampling capacitance: .
• MDAC variables.
II. PIPELINED ADC MODEL — Power: .
Illustrated in Fig. 1, the ADC model consists of four levels — Input/output loading: , .
of hierarchy: top-level, stage-level, circuit-level, and transistor- — Open-loop gain of the opamp: .
level. Top-level and stage-level model capture design tradeoffs — Input-referred noise power of the opamp: .
inherent in the pipelined ADC architecture, while circuit-level — Transconductance of the opamp: .
and transistor-level models deal with topology and technology — Input capacitance of the opamp: .
specific design tradeoffs. This section describes the top-level • Sub-ADC variables.
and stage-level ADC model. The circuit-level model as well — Power: .
as a brief introduction to the GP optimization are presented in — Input loading: .
Appendices A and B. — of input-referred offset: .
— Clock-to-data delay: .
A. Model Overview Note that the stage resolution is linked to the various stage
Shown in Fig. 2, top-level model consists of stages design parameters. For instance, the number of 1-bit flash slices
connected in cascade. Front-end stage, or stage, is a track- in the sub-ADC is given by . Also, ideal
and-hold (T/H) amplifier that samples the input and passes held stage gain is given by , where
analog output to the following ADC stages. Internal pipeline and are total capacitance of the two banks of sampling
stage, or th stage, resolves -bit from the analog input and capacitors in Fig. 3, respectively (e.g., ).
passes the residual analog information, or residue, to the th Since the total input capacitance is given by
stage. Backend stage is a special case of the stage which does , and can be chosen subject to and . Note
not pass residue to the following stage. that holds for the backend that does not need
Fig. 3 illustrates a generalized stage model as a single-ended redundancy.
representation. The stage model consists of two building blocks,
an MDAC and a sub-ADC. The topology of the MDAC fol- C. Equation-Based ADC Model
lows a standard flip-around switched-capacitor circuit using a • Total Resolution: Total resolution of the ADC, , must
high-gain opamp with negative feedback as a residue ampli- be same as the summation of the stage resolutions, i.e.,
fier [15]. The sub-ADC is a flash ADC consisting of slices,
each resolving 1-bit of digital information with corresponding
(1)
threshold value. The stage model adopts a standard 1-bit redun-
dancy in the sub-ADC of internal stages [16] in order to relax the
offset requirement in the sub-ADC. Note that T/H and backend where is the total number of the pipeline stages.
stage can be modeled as a special case of the internal stage. The Note that is a discrete variable, which is not found
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KIM et al.: MULTILEVEL POWER OPTIMIZATION OF PIPELINED A/D CONVERTERS 3

by solving a single optimization problem. Rather, we


have to choose the per-stage resolution before optimizing
the problem, e.g.,
or for , where is th
component of . Because optimizing the problem via GP
is very efficient, we can iteratively solve the problem for
various choices of that satisfies (1) to determine which
gives the best result.
• Total power dissipation: Total power of the ADC is the Fig. 4. (a) MDAC model in amplification phase. (b) Linear settling model of
summation of the stage powers, i.e., MDAC output.

(2)

• Input-referred signal-to-noise ratio : SNR is a ratio (6)


between signal power and the input-referred thermal noise
power, and needs to be lower-bounded to achieve a desired where is the effective feed-
noise performance. One can show that the SNR due to the back factor. Second, the dynamic error is a function of both
thermal noise in the pipelined ADC can be constrained as and , where the time constant can be approximately
expressed as [17]

(7)

where is the signal power of the sinusoid with where . Note that the total
peak-to-peak amplitude of , is input referred output capacitance of th stage, , is a lumped total ca-
thermal noise power of the stage and is a pacitance from both internal parasitic and following stage
given specification on the SNR. Since the stage adds noise loading, i.e.,
in the sampling and amplification phase, should
include two uncorrelated noise powers in both phases, i.e.,
(8)
(3) In a pipelined ADC design, in each stage is typi-
cally designed to be sufficiently smaller than the LSB of
remaining stage resolution to preserve the total linearity.
where is the Boltzman constant, is the absolute tem-
Since in (5) contains two components, we can as-
perature and is the time-constant of the MDAC, which
sign each contributing the same amount of error to the
shall be given by (7).
total gain error.1 Note that is also constrained by the
• Stage gain error: An MDAC in the amplification phase can
timing budgets in the amplification phase. Specifically, for
be modeled as a feedback amplifier shown in Fig. 4(a),
a given half-cycle period, non-overlapping clock period be-
and its dynamic response can be approximated as a first-
tween sampling and amplification phase, , and a
order linear settling when the slewing is negligible, which
timing delay in the sub-ADC to generate full logic level
is shown in the Fig. 4(b). While the ideal stage gain is
from sub-ADC input, , add to the total timing
given by , the actual stage gain deviates from
budget, i.e.,
due to the finite open-loop gain and time-constant of the
opamp. In a pipelined ADC that doesn’t incorporate some (9)
calibration techniques, stage gain error is the major source
of the ADC linearity degradation. One can show that the
• Sub-ADC offset: The input-referred offset of the sub-ADC
MDAC output voltage at the end of amplification phase can
needs to be less than allowable correction range in the
be modeled as
stage. Otherwise, output voltage of the stage exceeds the
(4) input range of the following stage, causing linearity error
that cannot be corrected by the digital redundancy. This
where is a static error, is a subtracted voltage in consideration can be constrained as
the MDAC, is a given time period for the settling, and
is the time constant of the MDAC. The total gain error (10)
from both nonzero and can be further simplified as
where is a yield parameter that is linked to the percentage
(5) of the sub-ADC whose offset lies beyond the allowable
We identify two sources of errors in (5). First, the static range in statistical sense.
error arises from the finite open-loop gain of opamp in 1From power dissipation perspective, it is not necessarily an optimal way to
the feedback configuration and can be expressed as [17] split the error evenly, and this is further discussed in Section IV.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS

III. DEVICE-TYPE AND SUPPLY COOPTIMIZATION


With the ADC model developed in Section II, this section
describes a design approach that jointly optimizes device-type
and supply voltages in an MDAC design to minimize the power
dissipation.
A. Device-Type Optimization
In an MDAC design that employs an opamp with negative
feedback as a residue amplifier, achieving high open-loop gain
is a common design goal to meet the required accuracy of
the stage. One straightforward way is to use non-minimum
channel length transistors in an opamp to enhance the intrinsic
gain , but it often leads to high power dissipation due
to the tight power-performance tradeoffs in an opamp design.
For instance, the tradeoff involved at the input transistor of an Fig. 5. Simulated intrinsic gain versus L of three types of NFET in 90-nm
opamp can be expressed as CMOS.

(11)
where is the cutoff frequency of the input tran-
sistor. Since increasing channel length of the transistor trades
for , (11) implies that the net benefit of increased
to the total gain error in (5) diminishes by the degraded feed-
back factor and time constant. Additionally, it has been rec-
Fig. 6. Tradeoff in selecting supply voltages and signal swing.
ognized that exercising the tradeoff between and by
the channel length is noticeably less efficient due to the adverse
effect of a pocket implant, a commonly-adopted channel engi-
neering method in sub-100-nm technology [18]. Therefore, an in an MDAC design. Note that while the advantage of choosing
attempt to improve by sizing the transistor with longer one device over the other may seem apparent at the device-level,
channel length is expected to incur excessively high power dis- the best combination of device types as well as the amount of net
sipation in nanometer-scale technologies. benefit in a particular circuit structure has to be carefully exam-
One way to relax the tight tradeoff is to leverage the multi- ined by cooptimizing the bias condition and sizing of individual
tude of devices offered in nanometer CMOS technology. While transistors.
this is only applicable method for the process that offers mul-
tiple device types, most of today’s standard foundry processes B. Supply Voltage Optimization
commonly offer a device pool with multiple threshold voltages In mixed-signal designs, supply voltages can often be design
and oxide thickness for designers to choose from. The primary variables when they are not constrained by system-level speci-
use of these devices is to help with digital designs to control fication. The impact of supply voltage to noise-limited analog
leakage and handle I/O voltages, but their differences in output circuit has been discussed considerably in previous literature
resistance can be exploited in analog designs. [9], [19]. Fundamentally, low supply voltage reduces the avail-
Fig. 5 shows a simulated intrinsic gain versus channel able headroom for signal swing, hence degrading the tradeoff
length of regular-threshold (CORE), high-threshold (HVT), between signal-to-noise ratio (SNR) and power dissipation.
and thick-oxide (I/O) type of NFET offered in a foundry 90-nm Fig. 6 illustrates a first-order tradeoff involved in the context
CMOS technology for an identical gate overdrive and head- of pipelined ADC. The SNR is related to the signal swing and
room condition. Evidently, a reduced sensitivity of to the sampling capacitor as
of CORE device results in marginally increasing intrinsic gain
when 0.2 m. However, it is worth noting that the incre- (12)
mental improvement of over varies substantially de-
pending on the type of device, particularly at longer channel
length. For instance, at 0.5 m, of I/O device is where is peak-to-peak signal swing and is sampling
more than 2 higher than that of the CORE device, with the capacitance. It follows from (12) that linearly increasing
gap reaching up to 5 at 1 m. The comparison shown in enables a quadratic reduction of for a fixed SNR require-
Fig. 5 suggests that when the highest is not needed, selecting ment. Using the gain-bandwidth product, or ,
I/O device instead of CORE device with same and overdrive as a measure of the speed, the transconductance and
leads to higher open-loop gain of the amplifier without com- correspondingly the bias current for a given overdrive voltage
promising , since is first-order independent to the oxide can be quadratically reduced for the same speed requirement.
thickness for identical overdrive and . This is in a sharp con- Therefore, having a large signal swing can potentially reduce
trast with merely increasing , which must compromise to the power dissipation of pipelined ADC.
improve . Therefore, we can anticipate that a judicious de- Increasing signal swing, however, imposes a tighter head-
vice-type selection should relax the power-performance tradeoff room constraint that exacerbates the open-loop gain of an
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KIM et al.: MULTILEVEL POWER OPTIMIZATION OF PIPELINED A/D CONVERTERS 5

TABLE I
DESIGN SPECIFICATION FOR T/H STAGE

500 ps is the non-overlapping period.

Fig. 7. (a) T/H stage model used in the analysis as a test-bench circuit. (b) stage model that can be optimized for a foundry 90-nm CMOS
Schematic of the telescopic opamp which is used as a stage amplifier.
process. More details on the optimization methodology as
well as the circuit-level model for telescopic opamp can be
found in Appendices A and B, respectively. Design variables
amplifier because the output resistance of transistor is in- include transistor-level variables such as width , length
versely proportional to . To accommodate larger signal , and bias condition of the transistors in a
swing without compromising the headroom constraint, one can telescopic opamp as well as supply voltages ,
increase analog supply while keeping the same gate signal swing , and sampling capacitance . The
overdrive, or , for all transistors. The net result is that model is optimized for minimum power dissipation for a set
analog power reduction is at best linearly proportion to the of design constraints listed in Table I, which is chosen for
signal swing. In addition, increasing is eventually limited the front-end stage of an 8-bit pipelined ADC. Note that the
by the device reliability. A common design constraint is that the tradeoff analysis demonstrated here is easily applicable to other
voltage across the gate oxide must be smaller than a designated amplifier topologies and design specification by modifying
limit set by the process during both the normal operation and equation-based circuit models.
power up-down sequence.2 Therefore, the net amount of power To find an optimal set of device types, a constrained but ex-
reduction from higher signal swing will depend highly on a haustive search has been performed. Since the telescopic opamp
circuit topology and process technology. consists of five transistors in each side and we consider three de-
Note that increasing signal swing has an impact on the power vice types (CORE, HVT, and I/O), the total number of possible
dissipation in the digital domain as well. In a typical pipelined device combination is . After optimizing the model for
ADC design, analog signal has to pass through the switches all 243 cases, the optimal device-type is chosen by picking the
whose allowable range depends on digital supply and one with the lowest power dissipation. Although this approach
threshold voltage . While the actual voltage range of the is by no means an efficient method, optimizing the stage model
signal is determined by a particular amplifier design, larger takes only 2.5 seconds in our linux machine running on Xeon
signal swing to the first order requires higher to increase 2.8 GHz CPU with 2-GB memory, making an exhaustive search
allowable signal range in the switches. Consequently, higher among hundreds of data sets a feasible method. The advantage
signal swing leads to a quadratic increase in the digital power of relying on such a numerical design framework is obvious; the
dissipation, degrading again the potential power reduction by proposed device selection would have required tremendous de-
increasing signal swing. sign effort if every 243 circuit designs were to be handcrafted,
Foregoing observations reveal that there may exist optimal let alone the optimality of each design.
and that achieve optimal balance between power Fig. 8 shows a histogram of the power dissipation collected
dissipation and SNR, which depend on a particular design spec- from the exhaustive search. There exist 144 feasible designs that
ification as well as circuit topology and process technology.3 meet given design specifications out of 243 combinations be-
In conjunction with the device type selection proposed in cause the required open-loop gain and time constant are not si-
Section III-A, a numerical design framework is necessary to multaneously met in all cases. The power distribution is widely
quantitatively explore the tradeoffs. spread from 0.61 to 4.2 mW, indicating that there is consider-
able impact of device-type selection on the power-performance
C. Tradeoff Analysis tradeoff of the T/H stage design.
To quantitatively explore the proposed design optimizations, Among 144 feasible solutions, Table II shows the design vari-
the stage model introduced in Section II has been extensively ables of the best device type as well as other design variables that
optimized for tradeoff analysis. A particular circuit topology result in the lowest power dissipation. The optimal design oc-
used here is a front-end T/H stage using telescopic opamp as curs when CORE device with close-to-minimum channel length
a residue amplifier as shown in Fig. 7. A generalized B-bit 0.12 m is used for the input transistor to simultane-
stage model previously shown in Fig. 3 is tailored for a T/H ously achieve high and , while long-channel I/O de-
vice 1.12 m in the cascoding transistor maximizes
2During the power-up sequence, one simple technique to relieve the device
the open-loop gain while meeting the phase margin require-
reliability concern is to set up the bias of the amplifier slightly earlier than ment. For the verification purpose, Table III shows a compar-
power-up of the main amplifier.
3Having separate supply voltages for analog and digital domain is not un-
ison between predicted performance metrics from the optimizer
common in mixed-mode circuit designs as separate supply helps isolate digital and simulated results, showing that there is a reasonably good
switching noise from analog supply. matching between them.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS

TABLE IV
TEN BEST DEVICE SELECTIONS

Fig. 8. Power histogram of track-and-hold stage from exhaustive search opti-


mizations.

TABLE II
OPTIMAL DESIGN VARIABLES FOR T/H DESIGN

Fig. 9. Optimal tradeoff between power dissipation of T/H and stage gain error.

TABLE V
OPTIMAL DEVICE SELECTION VERSUS GAIN ERROR SPEC

TABLE III
PREDICTED PERFORMANCE AND SIMULATED RESULT

gain error constraint and are listed in Table V; and 3) sub-op-


timal device selection in the Table II, which was found for an
8-bit equivalent gain error. Fig. 9 compares obtained power dis-
The multitude of histogram count in the first bin in sipation by optimizing the model. We first notice that a globally-
Fig. 8 indicates that a group of device-type selections achieves optimal case and a sub-optimal case almost overlap.
near-optimal result. Table IV lists the best-ten device combina- While the optimal device selection in Table V slightly changes
tions in the first bin. Emphasized in boldface, they commonly with gain error specifications, the comparison shows that the
choose CORE device for and I/O device for . The impli- difference is negligible if the device types of critical transistors,
cation is that device selections for transistors in the signal path and in this case, are judiciously chosen, agreeing with
have a more profound impact on the performance while the de- the observation made earlier in this section. A comparison with
vice types of other transistors are not as influential. Also, it is the case that uses only CORE device shows that the optimal
unlikely that the optimal device types in the signal path change device selection reduces the power dissipation by up to 60% or
over process corners since they are mainly circuit-topology de- gain error by 3 , which is a considerable improvement in the
pendent. This observation may be leveraged for larger problems power-performance tradeoff.
having more complicated circuit topologies to help reduce the Fig. 10 specifically explores the tradeoff involved in supply
search space. voltage and signal swing selection. Using the device types in
To quantitatively compare the amount of benefit by the pro- the Table II, it shows the optimized stage power dissipation
posed approach, we optimized the track-and-hold stage model versus for several selected . While the existence
while varying gain error constraints for three different scenarios: of optimal is easily identified ( 0.61 mW when
1) no device selection (CORE device for all transistors); 2) glob- 1.7 V and 1.21 V), it is also worth noting
ally-optimal device selection for corresponding gain error con- that using 1.2 V achieves the lowest the power dissi-
straints, which has been found by exhaustive searches for each pation. Since analog power dominates ( 540 W and
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KIM et al.: MULTILEVEL POWER OPTIMIZATION OF PIPELINED A/D CONVERTERS 7

brated pipelined ADC. A common wisdom has been to design


a stage accurately enough to preserve combined linearity of the
remaining stage resolutions. For example, one can design a stage
with a gain error less than 1/4 LSB of the remaining stage reso-
lution as described in Section II.
Two observations can be made. First, the required accuracy
diminishes along the pipeline, progressively relaxing the gain
error requirement. Together with the sampling capacitor scaling
[6], power dissipation of the first few stages should dominate the
total power dissipation in an optimized design. Second, given
the fact that the first few stages consume the majority of the
total power and their marginal power cost to reduce gain error
is higher than that of the remaining stages, relaxing their gain
error requirements would lead to substantial reduction in the
total power dissipation. From the linearity perspective, while the
relaxed gain error leads to an increased differential nonlinearity
(DNL), it is worth noting that integral nonlinearity (INL) is the
result of the cumulative contribution of DNL. In other words,
the gain errors of all stages, with diminishing weighting factor,
contribute to the INL degradation. Therefore, relaxing the gain
errors in the first few stages and tightening the gain errors in
the later stages can potentially save combined power dissipation
with INL performance being nearly unchanged.
The observations suggest that there can be a design optimiza-
tion that can judiciously balance the gain error contributions
from all stages while minimizing total power dissipation. We
refer to this as a gain error optimization.

Fig. 10. (a) Optimal power dissipation of T/H stage versus V for several B. Gain Error Model
V values. (b) Correspoding optimal peak-to-peak swing (V ).
To carry out proposed gain error optimization, the gain error
model compatible with GP optimization has been developed.
Because a compact equation that accurately links gain error to
70 W) in the T/H model for a given design speci-
the worst-case linearity degradation of an ADC is not straight-
fication, higher signal swing enabled by higher leads to
forward, we propose a roundabout metric by referring stage gain
a lower combined power dissipation. Note that these optimal
errors to the input of the converter, i.e.,
supply voltages, 1.73 V and 1.2 V, vary
depending on the design specifications and other stages in the
pipelined ADC. As shall be shown in Section V, 2.1 (13)
V is chosen as optimal analog supply when taking into account
all stages in the pipelined ADC.
where is a total resolution of ADC. Mathemat-
IV. GAIN ERROR OPTIMIZATION ically, (13) is an input-referred sum of th stage gain errors
Reusing the optimal design of a stage in cascade does not , normalized by the LSB of the converter.
accomplish a globally optimal design. Individually optimizing The correlation can be verified by behavioral simulations.
pipeline stage therefore has been of considerable interest in the Taking an 8-bit pipelined ADC converter as a test vehicle, stage
pipelined ADC design. Traditionally, this has meant optimizing gain errors less than 0.5 LSB of remaining stage resolution are
the stage noise contribution toward the input by optimizing randomly generated and INL has been found by behavioral sim-
the sampling capacitance and per-stage resolution while mini- ulations using MATLAB. Shown in Fig. 11 is a simulated worst-
mizing the total power dissipation. This work extends the scope case INL versus in (13) for 100 experiments,
by optimizing the nonlinearity contribution from stage gain where a linear correlation between and the can be
error. identified. Notice that (13) is not compatible with GP optimiza-
tion in its native form, but can be transformed to a compatible
A. Stage Gain Error Optimization form via numerical function fitting described in Appendix C.
The gain error of the stage shifts the residue output from
C. Tradeoff Analysis
its ideal value, and its cumulative effect over the stages turns
into a linearity degradation of the overall transfer character- This section quantitatively explores potential power saving by
istic. While the gain error can be calibrated by various tech- the gain error optimization in addition to optimizing per-stage
niques in both analog [12] and digital domain [11], [13], it re- resolution and noise contribution. The top-level ADC model in-
mains as a major source of linearity degradation in an uncali- troduced in Section II is used as a testbench circuit with cir-
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS

Fig. 12. Optimal power dissipation versus various bit-per-stage selections for
8-bit total resolution.

TABLE VII
Fig. 11. Worst-case INL versus 6 from behavioral simulations. POSSIBLE B FOR 8-bit RESOLUTION

TABLE VI
DESIGN SPECIFICATION FOR 8-bit 400 M ADC

without significantly demanding more power dissipation than


cuit-level model described in Appendix B. It has previously the optimal case.
been shown that the total power dissipation and input-referred Fig. 12 also shows how much power benefit can result from
SNR of such a top-level pipelined ADC model is compatible having individual supply domains for each stage as opposed to
with GP optimization [20]. Equation (13) is added to the ADC a shared and across all stages. When comparing
model in order to enable gain error optimization. The ADC the result between individual and shared supplies , the
model is optimized for minimum power dissipation with the de- sharing causes only 6% power penalty at the optimal case.
sign specifications shown in Table VI, where the specifications Therefore, the benefit is not considerable given the added
are chosen for a two-way interleaved 8-bit 400 MS/s ADC. Note design complexity.
that optimizing the top-level model allows us to explore not only With the optimal per-stage resolution of , we specifically
the optimal distributions of gain errors but optimal stage reso- explored the benefit of the proposed gain error optimization via
lution and sampling capacitance while individually optimizing comparative design optimizations. Two scenarios of design op-
the transistor-level device sizing and biasing. timization have been performed. The first case is when the ADC
Several design optimization have been performed. First, to model is optimized while uniformly preassigning stage gain
determine an optimal per-stage resolution, the model is iter- error to be 0.2 LSB of remaining stage resolution. The second
atively optimized for various choices of per-stage resolution case is when the stage gain error is optimally chosen by the pro-
whose sum equals to 8-bit, i.e., , where we consider posed approach. Shown in Table VIII, the result indicates that
either 1- or 2-bit for the internal stage resolution for simplicity. the error of the T/H and first stage in the proposed approach can
Twelve such candidates are listed in Table VII, where the th be relaxed by and , re-
component of is the resolution of th stage. Fig. 12 shows spectively, while achieving similar total linearity performance
the optimized power dissipation for the corresponding choice (simulated of 0.23LSB versus 0.3LSB) by assigning
of per-stage resolutions, where the number in the horizontal more stringent gain error in the rest of the stages. This optimal
axis is the index of vectors in Table VII. Several interpre- balancing leads to significant power saving compared to uni-
tations can be made from Fig. 12. First, the comparison indi- formly-assigned gain error case (31 mW versus 12.12 mW). In-
cates that a minimum power dissipation is obtained for tuitively, the net power reduction shown in Table VIII agrees
, which is the optimal case. Second, resolving with the observation made in Section IV-A. The power-perfor-
two-bits in the first stage in general leads to lower-power dissi- mance tradeoff in the T/H and first stage is very tight, therefore
pation (3–8, 10, 12 versus others in terms of index), which relaxing the tradeoff leads to substantial power reduction. On
agrees with a common design practice [21]. Third, however, the other hand, tightening the tradeoff in the rest of the stages
the bit distribution among the rest of the stages still has signif- incurs insignificant power cost.
icant impact on the power dissipation, causing up to 2 dif- The error partition between and in the Table VIII
ference (e.g., 6 versus 10) in the total power dissipation. Note also merits an attention. For most stages, the finite opamp gain
that one may opt for a non-optimal for other considerations contributes more error than the incomplete settling (e.g., 79%
such as layout or required design effort. For instance, choosing versus 21% in T/H). For the speed and resolution specification
can result in a more compact layout of our design, the marginal power cost to reduce settling time is
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KIM et al.: MULTILEVEL POWER OPTIMIZATION OF PIPELINED A/D CONVERTERS 9

TABLE VIII
COMPARISON FOR GAIN ERROR OPTIMIZATION

Fig. 14. Stage power and sampling capacitance scaling.

tightly designed in terms of gain error and noise, and stage-4


Fig. 13. Top-level ADC architecture. also dissipates as much power due to substantial input capaci-
tance arising from seven 1-bit flash slices in the 3-bit backend
flash converter. The capacitance scales up from T/H to stage-1
cheaper than an increase in the open-loop gain, which reflects because there is no gain in the track-and-hold stage therefore
the difficulty of achieving high open-loop gain in sub-microm- noise contributions from stage-1 and T/H are equally weighted.
eter technology. Given that the feedback factor in the T/H is close to 1, it can
still be designed with low power while driving a large capaci-
V. CHIP IMPLEMENTATION tive load from stage-1.
The proposed design optimizations have been applied to im-
plement a low-power 8-bit 320 MS/s pipelined ADC in 90-nm B. Stage-Level Design
CMOS technology. This section describes the architecture and A standard flip-around MDAC is used to implement all
circuit-level details of the design. pipeline stages. A telescopic opamp in Fig. 7(b) with optimized
device type in Table II is used as the stage amplifier. Chosen by
A. Architecture-Level Design the design optimization described in Section III-B, 2.1
The architecture of the ADC is shown in Fig. 13. The ADC V and 1.2 V with 0.8 V minimize the total
is organized in two slices, each consisting of an input T/H and power dissipation while meeting the given SNR requirement
pipelined ADC running at 160 Ms/s. The T/H is a standard for an 8-bit ADC. Compared to the analysis in Section III-C
flip-around architecture that employs a bootstrapped clock for that only considers T/H stage, optimal is higher (2.1
the tracking switch. The first stage in the ADC slice is a 2.8-bit V versus 1.73 V) while optimal swing is smaller (0.8
stage, followed by three 1.5-bit stages. The backend stage is V versus 1.23 V). The selection may look contradictory, but
a 3-bit flash converter. A 640 MHz input clock is twice-di- the 1st stage MDAC that resolves 2-bits puts more stringent
vided to generate a low-jitter 160 MHz clock, which subse- gain-error requirement than T/H due to the reduced ( 0.25 as
quently drives per-stage non-overlapping clock generators that opposed to 1 in T/H) feedback factor, favoring higher supply
are shared by both channels. Raw digital outputs from stage voltage and smaller signal swing to maximize open-loop gain
ADCs are combined by two levels of 2-to-1 multiplexers, gen- and minimize time constant.
erating eight output streams each at 640 Mb/s that are collected Note that while 2.1 V is higher than nominal supply voltage
by an off-chip logic analyzer for further analysis. of a 90-nm CMOS process, the design ensures that none of the
Shown in Fig. 14 is the stage power and sampling capacitor circuit components are stressed beyond its designated voltage
scaling adopted in the design. The largest power is assigned to limits under all operating conditions. Specifically, single-ended
stage-1 that resolves the two most significant bits, followed by voltage range at the output node of the telescopic opamp is be-
a rapid power and capacitor scaling from stage-1 to stage-2 to tween 0.8 and 1.2 V along with input common-mode voltage of
reduce capacitive load to stage-1. There is essentially no power 0.6 V. Therefore, top pMOS transistors ( and in Fig. 7)
scaling from stage-2 to stage-3 because these stages are more experience 1.3 V of combined headroom, or 0.65 V of
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS

Fig. 15. Dynamic comparator in the stage sub-ADC.

each, in a worst case, not jeopardizing the voltage stress limit.


In the case of stage-1, the simulated performance of the opamp
shows a dc gain of 62 dB and gain error of 0.6% for 208 fF Fig. 16. Die micro-photograph.
of total load capacitance, while dissipating 1.65 mW of static
power.
The sub-ADC in stage-1 comprises of six 1-bit flash slices,
each of which consists of a dynamic comparator preceded by a
switched-capacitor reference generator [22]. The dynamic com-
parator, shown in Fig. 15, adopts the switch in the signal
path to reduce static leakage power. Once the decision is made,
a DEC signal flips from low to high, disconnecting the current
path by closing to reduce the leakage flowing through
and .
Three subsequent stages (stages 2–4) are similarly realized as
1.5-bit stages using same circuit topologies but with individu-
ally optimized sizings. The backend stage is a 3-bit flash con-
verter consisting of seven 1-bit flash slices followed by a ther-
mometer-to-binary encoder. The backend flash converter em-
ploys an additional preamplifier between the reference gener- Fig. 17. Measured DNL and INL.
ator and comparator to reduce the input-referred offset, which
is not corrected by the digital redundancy built in the internal
is attributed to the nonlinearity in the front-end track-and-hold
stages.
amplifier. Fig. 19 shows SNDR for different sampling rates with
an 154.7 MHz sinusoidal input. The measured SNDR stays rel-
VI. EXPERIMENTAL RESULT atively constant up to 320 MHz and rolls off to 41.3 dB, which
The prototype ADC has been fabricated in a 90-nm eight- is below 7-bit of ENOB, at 360 MHz. While the design opti-
metal one-poly digital CMOS process. A micro-photograph of mization has been performed for target frequency of 400 MHz,
the implemented chip is shown in Fig. 16, where the ADC core the ADC model used in the optimization did not properly ac-
occupies a silicon area of 750 m 700 m. count parasitic capacitance in the signal path, resulting in mea-
Fig. 17 shows the DNL and INL obtained from a code den- sured ENOB of 7-bit up to 80% of the target speed. For future
sity test. Total 819000 data samples are collected for a low-fre- designs, the ADC model can be modified to include extracted
quency sinusoid input to generate the linearity profile. The mea- parasitic capacitance to yield more accurate performance pre-
sured DNL and INL are 0.45 0.35 LSB and 0.89 0.72 diction.
LSB, respectively. Fig. 20 shows the output spectrum when the input sinusoid is
The dynamic performance of the ADC has been measured at 154.7 MHz. The amplitude of input sinusoid is 0.3 dB below
by sweeping the frequency of an input sinusoid from near dc full-scale. The largest spur is the third-order distortion at a level
to the Nyquist frequency 160 MHz . Fig. 18 shows various of 57 dBc. The tone at 5.3 MHz results from
dynamic performance metrics over different input frequency at the gain mismatch between two interleaved paths at a level of
320 MS/s of sampling rate. The measured peak signal-to-noise 60 dBc. Input-referred offset mismatch also appears at
plus distortion ratio (SNDR) is 46 dB for a low-frequency si- at a level of 60 dBc.4 Note that all the spurs are well-below
nusoid input 3.76 MHz , which is equivalent to 7.35 b of the targeted 8-bit linearity.
effective number of bits (ENOB). Under the same condition, Table IX summarizes the achieved performance and mea-
the peak SNR and spurious-free dynamic range (SFDR) are sured power dissipation of the test chip. Measured total power
47 and 57 dB, respectively. At near-Nyquist input frequency dissipation of the chip is 12.77 mW, excluding digital output
154.7 MHz , the measured SNDR drops down to 45.7 dB, 4Input-referred offset mismatch between two interleaved path has been re-
or 7.3 b of ENOB. The degradation, which is relatively small, moved in the digital domain.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KIM et al.: MULTILEVEL POWER OPTIMIZATION OF PIPELINED A/D CONVERTERS 11

Fig. 18. Measured SNDR, SNR, and SFDR versus input frequency.
Fig. 21. Comparison of FOM with previously published 8-bit ADCs.

TABLE IX
SUMMARY OF PROTOTYPE ADC

Fig. 19. Measured SNDR versus various sampling rates.

8-bit pipelined converters with comparable sampling-speed.5 6


The comparison indicates that proposed power optimization
yielded an ADC with one of the best energy efficiency in its
class.

VII. CONCLUSION
This paper demonstrated an analytical design optimization
approach to reduce power dissipation of pipelined ADC. En-
abled by a model-based design optimization method via geo-
metric programming, presented quantitative tradeoff analysis as
well as measured result from a prototype ADC indicates that
substantial power reduction is possible by proposed device-type
and supply voltage selection at the circuit-level and optimal bal-
ancing of per-stage gain error at the architecture-level. Fabri-
Fig. 20. Frequency-domain response for near-Nyquist input.
cated in 90-nm CMOS technology, a prototype ADC optimized
by proposed design approaches achieves 7.30-bit of ENOB at
Nyquist input running at 320 Ms/s with best in-class power ef-
drivers and reference buffers. Out of this, 8.18 mW is con- ficiency of 253 fJ/conv.
sumed by the analog circuits, 0.99 mW by the sub-ADCs, Additional contribution of this paper is to demonstrate the
and 3.60 mW by the clock buffers and nonoverlapping clock viability of model-based design optimization for analog and
generators, respectively. To compare achieved power-per-
5We only compared 8-bit ADCs in 90-nm or older generations. It has been
formance metric, commonly referred figure of merit (FOM)
reported that an 8-bit ADCs in 65 nm such as [23] achieves similar conversion
for the ADC performance is calculated, which is defined as efficiency as this design.
. Fig. 21 compares FOM of the pro- 6Power dissipaton of the reference buffer is 5.6 mW, and it was not intended
totype chip (=253 fJ/conv-step) with that of recently published to be optimized by proposed design approaches.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS

mixed-signal circuits. While demonstrated for a pipelined ADC comparator. The circuit topology of the comparator is shown
design, the same approach can be applicable to a general class in Fig. 15.
of analog and mixed-signal building blocks. When combined A) Model Variables: The circuit-level model considers fol-
with analytical power optimization approach of digital circuits, lowing variables. Note that transistors in differential-symmetry
it will up the possibility of a truly global power optimization such as and are identically sized.
of complex mixed-signal VLSI driven from the top-level design • The width and length of all transistors: and
specification. .
• The and current of all transistor: and
APPENDIX A .
CIRCUIT OPTIMIZAITON VIA GP • The width of all switches: .7
Circuit optimization via geometric programming (GP) is a de- • The scaling factor for sub-ADC : .
sign methodology that has been gaining interest to optimize var- • The analog and digital supply voltages: and
ious mixed-signal circuit systems [14], [29]. The central idea is Here, is a total number of the switches in the stage, which
to develop a circuit model that captures design constraints for a depends on the resolution of the stage. Switches are either
predefined circuit topology, but the model has to be formulated pMOS or complementary depending on the signal range they
using special set of functions called posynomial and monomial. handle. The sub-ADC model is a simplified model that has a
Posynomial functions are real-valued functions of real, pos- single scaling variable, .
itive variable with nonnegative coefficients and any B) MDAC Model: The subscript in the transistor-level pa-
real exponents , i.e., rameters in this section refers to the transistor designator, e.g.,
is a gate-drain overlap capacitance of transistor in
Fig. 7(b).
(14)
• Power: is comprised of analog power dissipated in
opamp and digital power consumed in the switches
Posynomials are closed under addition, multiplication, and
nonnegative scaling. Monomial is a special case of a single-term (17)
posynomial, i.e.,
where is a summation of total active capacitance
(15) from all the switches.
• Terminal capacitance: The input capacitance at the sam-
Note that monomial is closed under division while posynomial pling phase is dominated by the sampling capacitor, i.e.,
is not.
GP is an optimization problem that has the following format: (18)

minimize The output capacitance due to self-loading can be ex-


subject to pressed as sum of parasitic capacitance, i.e.,

(19)
(16)
• Bias constraints: KCL dictates the relation between bias
where are posynomial functions and are monomial current of transistors, i.e.,
functions. Once a circuit model is formulated in the form of (3),
the model can be optimized for desired performance specifica- (20)
tions and an objective value over process corners to yield an
optimized circuit netlist. Limitations such as inability to handle Also, the summation of drain-to-source voltages of stacked
all functional forms can be alleviated by incorporating some it- transistors must be less than the supply voltage
erative or condensation methods. The important feature of GP
is that: 1) large problems can be very efficiently solved and 2) (21)
the obtained solution is globally optimal, enabling an efficient
exploration of globally-optimal design tradeoffs. Here, is the drain-to-source voltage with balanced in-
puts, or zero differential voltage.
APPENDIX B The transistors in the opamp must be in saturation region
CIRCUIT-LEVEL MODEL at the full-swing condition. This is particularly critical to
and which experience a large voltage excursion at
This appendix describes the circuit models used in the their drains. This condition translates into following design
pipeline stage in this paper. A particular circuit topology constraint
used in the MDAC is a fully-differential switched-capac-
itor circuit using telescopic opamp as a stage amplifier. The
opamp topology is shown in Fig. 7(b). The sub-ADC consists
(22)
of of 1-bit flash slices where each slice is a
switched-capacitor reference generator followed by a dynamic 7All switches use minimum channel length.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KIM et al.: MULTILEVEL POWER OPTIMIZATION OF PIPELINED A/D CONVERTERS 13

(34)

• Noise: Input-referred noise power of a telescopic opamp gain of differential mode to that of common-mode. We can
can be modeled as impose a lower bound as

(23) (31)

where is a thermal noise factor. Note the noise contri- or equivalently


butions from cascoding devices and are negligible
(32)
and therefore not included for simplicity.
• Open-loop gain: Ignoring body-effect, the open-loop gain
of telescopic opamp can be approximated by • Switches: On-resistance of the switches must not limit set-
tling response in the amplification phase. It can be com-
(24) monly constrained as

• Phase margin: Phase margin is defined as a total phase shift (33)


in the loop gain subtracted from at gain crossover fre-
quency . The gain crossover frequency of the telescopic where is the effective capacitance that a particular
opamp in Fig. 7(b) is approximately expressed as switch charges or discharge. In this model,
is used.
(25) C) Sub-ADC and Transistor-Level Model: An empirical
modeling method has been applied to create GP-compatible
We consider only one non-dominant pole arising from node sub-ADC model and transistor-level model in 90-nm CMOS
, which is given by process. Interested readers are referred to [30].

(26) APPENDIX C
POSYNOMIAL TRANSFORMATION OF THE INPUT-REFERRED
STAGE GAIN ERROR
Because the phase contribution by the dominant pole is
nearly 90 around unity gain bandwidth (assuming the One difficulty in incorporating (13) into the optimization
open-loop gain is larger than 10), the phase margin can then framework is that is not a posynomial function. It is
approximately be given by apparent from (5) that the second term in , or ,
does not conform to posynomial defined in (14). To address
(27) this issue, a posynomial approximation of is created by
a convex piece-wise linear fitting as8 shown in (34) at the top
of the page, where . The numerical error due
Here, we use an approximation of , which is
to this approximation is found to be 10.5% at the worst case
quite accurate for . By using (25) and (27), the
and 5% on average in the domain of and
phase margin can be constrained as
. Not only does this posynomial gain error
model enable optimal distribution of stage gain errors, it also
(28)
determines an optimal error partition between and
in a given stage.
• Slewing: We impose a lower-bound on the slew rate so that
the settling behavior is close to the ideal linear settling. The REFERENCES
slew rate can be constrained as [1] M. Horowitz, E. Alon, D. Patil, S. Naffziger, and R. Kumar, “Scaling,
power, and the future of CMOS,” in Proc. IEEE Int. Electron Devices
(29) Meet., 2005, pp. 7–15.
[2] K. Nagaraj, H. Fetterman, J. Anidjar, S. Lewis, and R. Renninger, “A
250-mW, 8-b, 52-Msamples/s parallel-pipelined A/D converter with re-
or equivalently duced number of amplifiers,” IEEE J. Solid-State Circuits, vol. 32, no.
3, pp. 312–320, Mar. 1997.
[3] B.-M. Min, P. Kim, F. I. Bowman, D. Boisvert, and A. Aude, “A
(30) 69-mW 10-bit 80-MSample/s pipelined CMOS ADC,” IEEE J.
Solid-State Circuits, vol. 38, no. 12, pp. 2031–2039, Dec. 2003.
• Common-mode rejection: Low-frequency common-mode 8Strictly speaking, (39) is a generalized posynomial, which still can be han-
rejection (CMRR) is defined here as the ratio between dc dled in GP.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS

[4] J. K. Fiorenza, T. Sepke, P. Holloway, C. G. Sodini, and H.-S. Lee, [26] H.-C. Kim, D.-K. Jeong, and W. Kim, “A 30 mW 8b 200 MS/s
“Comparator-based switched-capacitor circuits for scaled CMOS tech- pipelined CMOS ADC using a switched-opamp technique,” in Proc.
nologies,” IEEE J. Solid-State Circuits, vol. 41, no. 12, pp. 2658–2668, IEEE Int. Solid-State Circuits Conf., 2005, pp. 284–598.
Dec. 2006. [27] L. Brooks and H.-S. Lee, “A zero-crossing-based 8-bit 200 ms/s
[5] S. H. Lewis, “Optimizing the stage resolution in pipelined, multistage, pipelined ADC,” IEEE J. Solid-State Circuits, vol. 42, no. 12, pp.
analog-to-digital converters for video-rate applications,” IEEE Trans. 2677–2687, Dec. 2007.
Circuits Syst. II, Analog Digit. Signal Process., vol. 39, no. 8, pp. [28] Y. Shimizu, S. Murayama, K. Kudoh, and H. Yatsuda, “A split-load
516–523, Aug. 1992. interpolation-amplifier-array 300 ms/s 8b subranging ADC in 90 nm
[6] D. Cline and P. Gray, “A power optimized 13-b 5 m samples/s CMOS,” in Proc. IEEE Int. Solid-State Circuits Conf., Feb. 2008, pp.
pipelined analog-to-digital converter in 1.2 m CMOS,” IEEE J. 552–553.
Solid-State Circuits, vol. 31, no. 3, pp. 294–303, Mar. 1996. [29] S. Boyd, S.-J. Kim, and S. Mohan, “Geometric programming and its ap-
[7] J. Goes, J. Vital, and J. Franca, “Systematic design for optimization plications to EDA problems,” Proc. Des., Autom. Test Eur., Nov. 2004.
of high-speed self-calibrated pipelined A/D converters,” IEEE Trans. [30] J. Kim, “Multi-level design optimizations of pipelined A/D converter,”
Circuits Syst. II, Analog Digit. Signal Process., vol. 45, no. 12, pp. Ph.D. degree, Dept. Elect. Eng., Univ. California, Los Angeles, 2008.
1513–1526, Dec. 1998.
[8] P. Kwok and H. Luong, “Power optimization for pipeline analog-to-
digital converters,” IEEE Trans. Circuits Syst. II, Analog Digit. Signal Jintae Kim (S’02–M’08) was born in Seoul, Korea.
Process., vol. 46, no. 5, pp. 549–553, May 1999. He received the B.S. degree in electrical engineering
[9] Y. Chiu, P. Gray, and B. Nikolic, “A 14-b 12-MS/s CMOS pipeline from Seoul National University, Seoul, Korea, in
ADC with over 100-dB SFDR,” IEEE J. Solid-State Circuits, vol. 39, 1997, and the M.S. and Ph.D. degrees in electrical
no. 12, pp. 2139–2151, Dec. 2004. engineering from University of California, Los
[10] Y.-T. Chien, D. Chen, J.-H. Lou, G.-K. Ma, R. Rutenbar, and T. Angeles, in 2004 and 2008, respectively.
Mukherjee, “Designer-driven topology optimization for pipelined From 1997 to 2001, he was with Xeline, Seoul,
analog to digital converters,” Proc. Des., Autom. Test Eur., pp. Korea, designing baseband ICs for high-speed power
279–280, 2005. line communication. During the summer of 2003 and
[11] A. N. Karnicolas, H.-S. Lee, and K. L. Barcrania, “A 15-b 1-Msample/s 2004, he was with Barcelona Design, Sunnyvale, CA,
digitally self-calibrated pipeline ADC,” IEEE J. Solid-State Circuits, developing CAD algorithms for analog circuit opti-
vol. 28, no. 12, pp. 1207–1215, Dec. 1993. mization. Since 2008, he has been with the Mixed-Signal Electronics Depart-
[12] J. Ming and S. H. Lewis, “An 8-b 80-msample/s pipelined analog-to- ment, Agilent Technologies, Santa Clara, CA, as a member of technical staff.
digital converter with background calibration,” IEEE J. Solid-State Cir- His research interests include the design and CAD methodologies of high per-
cuits, vol. 36, no. 10, pp. 1489–1497, Oct. 2001. formance mixed-signal circuits in CMOS.
[13] B. Murmann and B. E. Boser, “A 12-bit 75-MS/s pipelined ADC using Dr. Kim was a recipient of the IEEE Solid-State Circuits Predoctoral Fellow-
open-loop residue amplification,” IEEE J. Solid-State Circuits, vol. 38, ship in 2007.
no. 12, pp. 2040–2050, Dec. 2003.
[14] M. Hershenson, S. P. Boyd, and T. H. Lee, “Optimal design of a CMOS
opamp via geometric programming,” IEEE Trans. Comput.-Aided Des.
Integr. Circuits Syst., vol. 20, no. 1, pp. 1–21, Jan. 2001.
[15] B. Song, M. F. Tompsett, and K. R. Lakshmikumar, “A 12-bit Sotirios Limotyrakis (S’91–M’05) was born in
1-Msample/s capacitor error-averaging pipelined A/D converter,” Athens, Greece, in 1971. He received the B.S.
IEEE J. Solid-State Circuits, vol. 23, no. 12, pp. 1324–1333, Dec. degree from the National Technical University of
1988. Athens, Athens, Greece, in 1995 and the M.S. and
[16] S. Lewis, H. Fetterman, J. Gross, R. Ramachandran, and T. Ph.D. degrees from Stanford University, Stanford,
Viswanathan, “A 10-b 20-msample/s analog-to-digital converter,” CA, in 1997 and 2005, respectively, all in electrical
IEEE J. Solid-State Circuits, vol. 27, no. 3, pp. 351–358, Mar. 1992. engineering.
[17] B. Razavi, Design of Analog CMOS Integrated Circuits. New York: In the Summer of 1993, he was with KDDI
McGraw Hill, 2000. Corporation, Saitama R&D Labs, Saitama, Japan,
[18] D. Buss, B. Evans, J. Bellay, W. Krenik, B. Haroun, D. Leipold, K. working on the design of communication protocols.
Maggio, J.-Y. Yang, and T. Moise, “SOC CMOS technology for per- During the Summers of 1996 and 1997, he was with
sonal internet products,” IEEE Trans. Electron Devices, vol. 50, no. 3, the RF Design Group, Texas Instruments, Inc., R&D Center, Dallas, TX, where
pp. 546–556, Mar. 2003. he focused on LNA, RF oscillator design, and GSM transmit architectures. In
[19] A.-J. Annema, B. Nauta, R. van Langevelde, and H. Tuinhout, “Analog November 2004, he joined Atheros Communications, Inc., Santa Clara, CA,
circuits in ultra-deep-submicron CMOS,” IEEE J. Solid-State Circuits, where he is now a Senior Member of the technical staff. His current research
vol. 40, no. 1, pp. 132–143, Jan. 2005. interests include the design of mixed-signal and RF circuits for low-power data
[20] M. Hershensons, “Design of pipeline analog-to-digital converters via conversion and broadband communications.
geometric programming,” in Proc. ICCAD, San Jose, CA, Nov. 2002, Dr. Limotyrakis was a recipient of the W. Burgess Dempster Memorial Fel-
pp. 317–324. lowship at Stanford University in 1995, the 2004 Analog Devices Outstanding
[21] W. Yang, D. Kelly, I. Mehr, M. T. Sayuk, and L. Singer, “A 3-V Student Designer Award, and the corecipient of the 2004 IEEE Beatrice Winner
340-mW 14-b 75-Msample/s CMOS ADC with 85-dB SFDR at Award for Editorial Excellence.
nyquist input,” IEEE J. Solid-State Circuits, vol. 36, no. 12, pp.
1931–1936, Dec. 2001.
[22] S. Limotyrakis, S. D. Kulchycki, D. K. Su, and B. A. Wooley, “A
150-MS/s 8-b 71-mW CMOS time-interleaved ADC,” IEEE J. Solid- Chih-Kong Ken Yang (S’94–M’98–SM’07) was
State Circuits, vol. 40, no. 5, pp. 1057–1067, May 2005. born in Taipei, Taiwan. He received the B.S. and
[23] W.-H. Tu and T.-H. Kang, “A 1.2 v 30 mW 8b 800 MS/s time-inter- M.S. degrees and the Ph.D. degree from Stanford
leaved ADC in 65 nm CMOS,” in Symp. VLSI Circuits Dig. Techn. University, Stanford, CA, in 1992, 1992, and 1998,
Papers, 2008, pp. 72–73. respectively, all in electrical engineering.
[24] P. Wu, V.-L. Cheung, and H. Luong, “A 1-V 100-MS/s 8-bit CMOS He joined the University of California at Los
switched-opamp pipelined ADC using loading-free architecture,” Angeles as an Assistant Professor in 1999 and has
IEEE J. Solid-State Circuits, vol. 42, no. 4, pp. 730–738, Apr. 2006. been a Professor since 2009. His current research
[25] J. Mulder, C. Ward, C.-H. Lin, D. Kruse, M. Westra, J. R. Lugthart, E. interests are high-performance mixed-mode circuit
Arslan, R. van de Plassche, K. Bult, and F. van der Goes, “A 21 mW design for VLSI systems such as clock generation,
8b 125MS/s ADC occupying 0.09 mm in 0.13 m CMOS,” in Proc. highperformance signaling, low-power digital func-
IEEE Int. Solid-State Circuits Conf., Feb. 2004, pp. 260–261. tional blocks, and analog-to-digital conversion

Potrebbero piacerti anche